Source-linked AI summary
Improving automated multiple sclerosis lesion segmentation with a cascaded 3D convolutional neural network approach
Sergi Valverde, Mariano Cabezas, Eloy Roura, Sandra González-Villà, Deborah Pareto, Joan-Carles Vilanova, LLuís Ramió-Torrentà, Àlex Rovira, Arnau Oliver, Xavier Lladó
TL;DR
Manual MS lesion annotation is time-consuming and variable, motivating automated WM lesion segmentation. The paper uses two cascaded 3D patch-wise CNNs, with the first finding candidates and the second reducing misclassified voxels. It outperformed participant methods on MICCAI2008 and correlated highly with expected lesion volume on clinical datasets.
Problem
Manual MS lesion annotation is time-consuming and prone to inter-observer variability, while labeled training data can be limited.
Method
The method uses two independently trained 3D patch-wise CNNs: CNN1 finds candidate lesion voxels and CNN2 reduces misclassified candidates.
Results
The method outperformed participant methods in the MICCAI2008 challenge and achieved r ≥0.97 correlation with expected lesion volume on clinical MS datasets.
Takeaways & Limitations
The cascade provides a validated automated WM lesion segmentation approach across public and private MS datasets and varied training conditions.
Takeaways & Limitations
The CNN must be trained and tested separately for each evaluated dataset, which is time-consuming and may require more expertise.
Abstract
from arXiv · showhide
In this paper, we present a novel automated method for White Matter (WM) lesion segmentation of Multiple Sclerosis (MS) patient images. Our approach is based on a cascade of two 3D patch-wise convolutional neural networks (CNN). The first network is trained to be more sensitive revealing possible candidate lesion voxels while the second network is trained to reduce the number of misclassified voxels coming from the first network. This cascaded CNN architecture tends to learn well from small sets of training data, which can be very interesting in practice, given the difficulty to obtain manual label annotations and the large amount of available unlabeled Magnetic Resonance Imaging (MRI) data. We evaluate the accuracy of the proposed method on the public MS lesion segmentation challenge MICCAI2008 dataset, comparing it with respect to other state-of-the-art MS lesion segmentation tools. Furthermore, the proposed method is also evaluated on two private MS clinical datasets, where the performance of our method is also compared with different recent public available state-of-the-art MS lesion segmentation methods. At the time of writing this paper, our method is the best ranked approach on the MICCAI2008 challenge, outperforming the rest of 60 participant methods when using all the available input modalities (T1-w, T2-w and FLAIR), while still in the top-rank (3rd position) when using only T1-w and FLAIR modalities. On clinical MS data, our approach exhibits a significant increase in the accuracy segmenting of WM lesions when compared with the rest of evaluated methods, highly correlating ($r \ge 0.97$) also with the expected lesion volume.
1. Introduction
MS lesion segmentation is clinically relevant but manual annotation is time-consuming and variable. The paper introduces a cascaded 3D CNN pipeline for automated WM lesion segmentation and evaluates it against existing methods.
- MRI is widely used to diagnose and monitor MS because it reveals WM lesions over time and space.
- Manual lesion annotation is time-consuming and prone to inter-observer variability, motivating automated segmentation methods.
- Existing MS lesion segmentation methods include supervised, ensemble, non-local-means, k-nearest-neighbor, probabilistic, and thresholding-based approaches.
- Only a few CNN methods had been introduced for WM lesion segmentation in MS patients.
- The paper proposes a cascade of two convolutional neural networks using 3D neighboring patches from multiple input modalities.
- The authors describe this as the first cascaded 3D CNN approach proposed for MS lesion segmentation.
2. Methods
The method uses multi-channel 3D MRI patches and two independently trained CNNs to address limited labeled data and severe lesion-voxel imbalance. CNN1 identifies candidate lesions, while CNN2 focuses on difficult false positives before final thresholding and region filtering.
- 2.1. Input features: Multi-channel 3D patches combine available MRI sequences and provide spatial contextual information for WM lesion segmentation.
- 2.1. Input features: Input patches are normalized, centered on voxels of interest, and represented as P = [n×c×p×p×p], where n is training voxels and c is input modalities.
- 2.1. Input features: The patch size is empirically set to p = 11 after testing sizes 9, 11, 13, and 15.
- 2.2. Cascade based training: Because lesion voxels comprise only about 1.5% of total brain volume in a 20 ml lesion example, the method addresses severe class imbalance.
- 2.2. Cascade based training: CNN1 is trained on all positive patches and an equal number of randomly sampled negative patches after removing low-intensity FLAIR candidates.
- 2.2. Cascade based training: CNN2 is trained from scratch on positive patches and negatives misclassified by CNN1, targeting difficult candidate voxels.
- 2.3. CNN architecture: The two networks use the same 7-layer architecture, with each network trained independently and without parameter sharing.
- 2.7. Implementation: At testing, CNN1 discards low-probability voxels, CNN2 produces the final probabilistic mask, and thresholding plus minimum-size filtering yields binary regions.
3.1. MICCAI 2008 MS lesion segmentation
The MICCAI 2008 evaluation used 45 scans, blind testing, and metrics measuring lesion-volume error, sensitivity, false-positive rate, and an overall score. The proposed cascaded CNN ranked first with all modalities and remained among the top three using only T1-w and FLAIR, while combining high sensitivity with lower false-positive bounds.
- Dataset and evaluation: The challenge comprised 45 scans: 20 training cases with expert annotations and 25 testing cases without released segmentations.Images included T1-w, T2-w, and FLAIR modalities at 0.5 mm isotropic resolution.
- Dataset and evaluation: Evaluation was blind, comparing submitted masks for 25 testing cases with manual annotations from UNC and CHB raters.The metrics were lesion-volume error, true-positive rate, false-positive rate, and a combined ranking score.
- Experimental setup: Two pipelines were trained using either all modalities or only T1-w and FLAIR, based on the 20 manually annotated training images.The training database contained 800,000 balanced patches, with 25% reserved for validation and early stopping used during training.
- Results: The all-modality pipeline ranked first among participant strategies, while the T1-w-plus-FLAIR pipeline ranked in the top three.The reduced-modality pipeline was outperformed only by Jesson and Arbel (2015).
- Results: The method achieved the lowest absolute lesion-volume difference for both reported rater configurations and maintained high sensitivity with lower false-positive bounds than competing high-sensitivity methods.Jerman et al. (2016) achieved higher sensitivity, but methods with high sensitivity, including Jerman et al. and Brosch et al., also tended to increase false positives.
3.2. Clinical MS dataset
The clinical evaluation used two MS datasets with differing MRI modalities and compared the cascaded CNN against established pipelines under two training procedures. The proposed method outperformed competing methods across detection, precision, segmentation, and lesion-volume estimation.
- Clinical datasets: 60 patients were divided into MS1 with T1-w, FLAIR, and PD modalities and MS2 with only T1-w and FLAIR.MS1 contained 35 subjects, while MS2 contained 25 patients because PD images were unavailable for those subjects.
- Evaluation: The evaluation complemented Dice overlap with lesion-detection, false-positive, and precision metrics, including TPR, FPR, and PPV.DSC measured agreement between manual annotations and output masks, while PPV measured precision against lesion-region candidates.
- Training procedures: Two training procedures were tested: leave-one-out cross-validation and training on one clinical dataset followed by testing on the other.The cross-dataset procedure inverted the training and testing datasets to evaluate generalization in both directions.
- Results: The cascaded CNN clearly outperformed SLS and LST across detection, false-positive control, precision, and segmentation metrics, with similar performance across datasets and training procedures.The comparison used pipelines trained with either leave-one-out experiments or different training-testing datasets.
- Results: Adding the second network increased sensitivity, reduced false-negative rates, and significantly improved DSC overlaps by reducing misclassified voxels.This comparison was made against the same CNN architecture without cascading.
- Results: r ≥ 0.97 in all cases, with CNN estimates closer to expected lesion volumes and narrower confidence intervals, especially at higher lesion loads.The proposed methods showed distinctively lower confidence intervals in both clinical datasets.
4. Discussion
The proposed cascaded CNN improves MS white-matter lesion segmentation by combining a sensitivity-oriented first network with a second network that reduces misclassified voxels. Across clinical and benchmark evaluations, it improves lesion detection and volume agreement, while requiring dataset-specific training and testing.
- Architecture: The cascade uses two identical CNNs: the first finds candidate lesion voxels, while the second reduces misclassified voxels from that candidate set.This design separates sensitive candidate detection from more specific discrimination.
- Volume agreement: r ≥0.97: CNN architectures showed the strongest Pearson correlations between estimated and manual lesion volumes across the clinical datasets.Distances between computed and expected lesion volumes were generally lower for CNN methods than for the other evaluated pipelines.
- Segmentation performance: The method maintains high lesion sensitivity while keeping false positives reasonably low, addressing the usual trade-off between sensitivity and false-positive control.Its average deviation from manual lesion volume was lowest, especially for images with high lesion size.
- Training considerations: Independent training of the two smaller networks reduces overfitting risk and makes the architecture suitable for datasets without massive numbers of training samples.The approach uses fewer parameters to optimize at each stage than a single deeper network.
- Limitations: The method must be trained and tested separately for each dataset, and learned representations may depend strongly on acquisition protocols and image domain.The authors therefore suggest using it only within the same image domain.
5. Conclusions
The paper presents a cascaded CNN method for automated MS white-matter lesion segmentation and evaluates it on public and private datasets. It outperforms participant methods on the MICCAI2008 benchmark and shows higher sensitivity with reasonable false-positive control and lower lesion-volume deviation.
- Conclusions: The method consistently performs well on both public and private MS datasets, outperforming the other participant methods in the MICCAI2008 challenge.The MICCAI2008 challenge is described as a benchmark for proposed strategies.
- Conclusions: The approach significantly increases lesion sensitivity while maintaining a reasonably low number of false positives.It also yields lower deviation between expected and manual lesion volumes across modalities and datasets.
- Practical relevance: The cascaded CNN tends to learn well from small datasets, which is relevant when manual annotations are difficult to obtain and unlabeled MRI data are abundant.The authors present this as a practical property of the architecture.
- Conclusions: The results bring the CNN architecture closer to human expert inter-rater variability, although further research is still needed.The authors describe the method as a potentially valid alternative for automated white-matter lesion segmentation.